C++ API 'nn::Sequential' has inconsistent behavior with python conterpart

[FluorineDog]

🐛 Bug

To Reproduce

C++ version

class FakeNet : public torch::nn::Module {
  public:
    FakeNet() {
        id = register_module("id", nn::Sequential());
    }
    torch::Tensor forward(torch::Tensor x) {
        // Runtime Error: Cannot call forward() on an empty Sequential
        return id->forward(x); 
    }
  private:
    nn::Sequential id;
}; 

Python Version

class FakeNet(nn.Module):
    def __init___(self):
        self.id = nn.Sequential() 
    def forward(self, x):
        x = id(x) # OK
        return x

nn::Sequential doesn't overload operator() like other functions (e.g., nn::Conv2d, nn::BatchNorm), while in its forward function, it has an AT_CHECK on is_empty() == false, which makes using nn::Sequential as identity function impossible.

Suggestion

The C++ version should implement nn::Sequential::operator()(...), which should silently return x when is_empty() == true.

Environment

• PyTorch Version (e.g., 1.0.1):

[ailzhang]

Yea our cpp API is not consistent with python api on this. Note that your python example was actually calling python id() instead of self.id().
A proper python example is:

import torch
import pdb
import torch.nn as nn


class FakeNet(nn.Module):
    def __init__(self):
        super(FakeNet, self).__init__()
        # self.conv = nn.Conv2d(1, 1, 1)
        self.conv = nn.Sequential()

    def forward(self, x):
        y = self.conv(x)
        return y


input = torch.rand(1, 1, 2, 2)
pdb.set_trace()
m = FakeNet()
print(m(input))

We should match the behavior I think.

[ailzhang]

We're happy to accept an PR for it, would you minding sending one? @FluorineDog

[FluorineDog]

Gladly, I will have a look tomorrow, but I have to go to bed now :)
BTW, I'm working on a C++ demo on ResNet50, currently runnable, but it still needs to be refined. I will send a PR when I'm done.
Cheers.